projects/linux·blob v5.2 · refs

2019 lines · 53.8 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright(C) 2005-2006, Thomas Gleixner <[email protected]>
4 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
5 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner
6 *
7 * High-resolution kernel timers
8 *
9 * In contrast to the low-resolution timeout API, aka timer wheel,
10 * hrtimers provide finer resolution and accuracy depending on system
11 * configuration and capabilities.
12 *
13 * Started by: Thomas Gleixner and Ingo Molnar
14 *
15 * Credits:
16 * Based on the original timer wheel code
17 *
18 * Help, testing, suggestions, bugfixes, improvements were
19 * provided by:
20 *
21 * George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
22 * et. al.
23 */
24
25#include <linux/cpu.h>
26#include <linux/export.h>
27#include <linux/percpu.h>
28#include <linux/hrtimer.h>
29#include <linux/notifier.h>
30#include <linux/syscalls.h>
31#include <linux/interrupt.h>
32#include <linux/tick.h>
33#include <linux/seq_file.h>
34#include <linux/err.h>
35#include <linux/debugobjects.h>
36#include <linux/sched/signal.h>
37#include <linux/sched/sysctl.h>
38#include <linux/sched/rt.h>
39#include <linux/sched/deadline.h>
40#include <linux/sched/nohz.h>
41#include <linux/sched/debug.h>
42#include <linux/timer.h>
43#include <linux/freezer.h>
44#include <linux/compat.h>
45
46#include <linux/uaccess.h>
47
48#include <trace/events/timer.h>
49
50#include "tick-internal.h"
51
52/*
53 * Masks for selecting the soft and hard context timers from
54 * cpu_base->active
55 */
56#define MASK_SHIFT (HRTIMER_BASE_MONOTONIC_SOFT)
57#define HRTIMER_ACTIVE_HARD ((1U << MASK_SHIFT) - 1)
58#define HRTIMER_ACTIVE_SOFT (HRTIMER_ACTIVE_HARD << MASK_SHIFT)
59#define HRTIMER_ACTIVE_ALL (HRTIMER_ACTIVE_SOFT | HRTIMER_ACTIVE_HARD)
60
61/*
62 * The timer bases:
63 *
64 * There are more clockids than hrtimer bases. Thus, we index
65 * into the timer bases by the hrtimer_base_type enum. When trying
66 * to reach a base using a clockid, hrtimer_clockid_to_base()
67 * is used to convert from clockid to the proper hrtimer_base_type.
68 */
69DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
70{
71 .lock = __RAW_SPIN_LOCK_UNLOCKED(hrtimer_bases.lock),
72 .clock_base =
73 {
74 {
75 .index = HRTIMER_BASE_MONOTONIC,
76 .clockid = CLOCK_MONOTONIC,
77 .get_time = &ktime_get,
78 },
79 {
80 .index = HRTIMER_BASE_REALTIME,
81 .clockid = CLOCK_REALTIME,
82 .get_time = &ktime_get_real,
83 },
84 {
85 .index = HRTIMER_BASE_BOOTTIME,
86 .clockid = CLOCK_BOOTTIME,
87 .get_time = &ktime_get_boottime,
88 },
89 {
90 .index = HRTIMER_BASE_TAI,
91 .clockid = CLOCK_TAI,
92 .get_time = &ktime_get_clocktai,
93 },
94 {
95 .index = HRTIMER_BASE_MONOTONIC_SOFT,
96 .clockid = CLOCK_MONOTONIC,
97 .get_time = &ktime_get,
98 },
99 {
100 .index = HRTIMER_BASE_REALTIME_SOFT,
101 .clockid = CLOCK_REALTIME,
102 .get_time = &ktime_get_real,
103 },
104 {
105 .index = HRTIMER_BASE_BOOTTIME_SOFT,
106 .clockid = CLOCK_BOOTTIME,
107 .get_time = &ktime_get_boottime,
108 },
109 {
110 .index = HRTIMER_BASE_TAI_SOFT,
111 .clockid = CLOCK_TAI,
112 .get_time = &ktime_get_clocktai,
113 },
114 }
115};
116
117static const int hrtimer_clock_to_base_table[MAX_CLOCKS] = {
118 /* Make sure we catch unsupported clockids */
119 [0 ... MAX_CLOCKS - 1] = HRTIMER_MAX_CLOCK_BASES,
120
121 [CLOCK_REALTIME] = HRTIMER_BASE_REALTIME,
122 [CLOCK_MONOTONIC] = HRTIMER_BASE_MONOTONIC,
123 [CLOCK_BOOTTIME] = HRTIMER_BASE_BOOTTIME,
124 [CLOCK_TAI] = HRTIMER_BASE_TAI,
125};
126
127/*
128 * Functions and macros which are different for UP/SMP systems are kept in a
129 * single place
130 */
131#ifdef CONFIG_SMP
132
133/*
134 * We require the migration_base for lock_hrtimer_base()/switch_hrtimer_base()
135 * such that hrtimer_callback_running() can unconditionally dereference
136 * timer->base->cpu_base
137 */
138static struct hrtimer_cpu_base migration_cpu_base = {
139 .clock_base = { { .cpu_base = &migration_cpu_base, }, },
140};
141
142#define migration_base migration_cpu_base.clock_base[0]
143
144/*
145 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
146 * means that all timers which are tied to this base via timer->base are
147 * locked, and the base itself is locked too.
148 *
149 * So __run_timers/migrate_timers can safely modify all timers which could
150 * be found on the lists/queues.
151 *
152 * When the timer's base is locked, and the timer removed from list, it is
153 * possible to set timer->base = &migration_base and drop the lock: the timer
154 * remains locked.
155 */
156static
157struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
158 unsigned long *flags)
159{
160 struct hrtimer_clock_base *base;
161
162 for (;;) {
163 base = timer->base;
164 if (likely(base != &migration_base)) {
165 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
166 if (likely(base == timer->base))
167 return base;
168 /* The timer has migrated to another CPU: */
169 raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
170 }
171 cpu_relax();
172 }
173}
174
175/*
176 * We do not migrate the timer when it is expiring before the next
177 * event on the target cpu. When high resolution is enabled, we cannot
178 * reprogram the target cpu hardware and we would cause it to fire
179 * late. To keep it simple, we handle the high resolution enabled and
180 * disabled case similar.
181 *
182 * Called with cpu_base->lock of target cpu held.
183 */
184static int
185hrtimer_check_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base)
186{
187 ktime_t expires;
188
189 expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
190 return expires < new_base->cpu_base->expires_next;
191}
192
193static inline
194struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base,
195 int pinned)
196{
197#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
198 if (static_branch_likely(&timers_migration_enabled) && !pinned)
199 return &per_cpu(hrtimer_bases, get_nohz_timer_target());
200#endif
201 return base;
202}
203
204/*
205 * We switch the timer base to a power-optimized selected CPU target,
206 * if:
207 * - NO_HZ_COMMON is enabled
208 * - timer migration is enabled
209 * - the timer callback is not running
210 * - the timer is not the first expiring timer on the new target
211 *
212 * If one of the above requirements is not fulfilled we move the timer
213 * to the current CPU or leave it on the previously assigned CPU if
214 * the timer callback is currently running.
215 */
216static inline struct hrtimer_clock_base *
217switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
218 int pinned)
219{
220 struct hrtimer_cpu_base *new_cpu_base, *this_cpu_base;
221 struct hrtimer_clock_base *new_base;
222 int basenum = base->index;
223
224 this_cpu_base = this_cpu_ptr(&hrtimer_bases);
225 new_cpu_base = get_target_base(this_cpu_base, pinned);
226again:
227 new_base = &new_cpu_base->clock_base[basenum];
228
229 if (base != new_base) {
230 /*
231 * We are trying to move timer to new_base.
232 * However we can't change timer's base while it is running,
233 * so we keep it on the same CPU. No hassle vs. reprogramming
234 * the event source in the high resolution case. The softirq
235 * code will take care of this when the timer function has
236 * completed. There is no conflict as we hold the lock until
237 * the timer is enqueued.
238 */
239 if (unlikely(hrtimer_callback_running(timer)))
240 return base;
241
242 /* See the comment in lock_hrtimer_base() */
243 timer->base = &migration_base;
244 raw_spin_unlock(&base->cpu_base->lock);
245 raw_spin_lock(&new_base->cpu_base->lock);
246
247 if (new_cpu_base != this_cpu_base &&
248 hrtimer_check_target(timer, new_base)) {
249 raw_spin_unlock(&new_base->cpu_base->lock);
250 raw_spin_lock(&base->cpu_base->lock);
251 new_cpu_base = this_cpu_base;
252 timer->base = base;
253 goto again;
254 }
255 timer->base = new_base;
256 } else {
257 if (new_cpu_base != this_cpu_base &&
258 hrtimer_check_target(timer, new_base)) {
259 new_cpu_base = this_cpu_base;
260 goto again;
261 }
262 }
263 return new_base;
264}
265
266#else /* CONFIG_SMP */
267
268static inline struct hrtimer_clock_base *
269lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
270{
271 struct hrtimer_clock_base *base = timer->base;
272
273 raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
274
275 return base;
276}
277
278# define switch_hrtimer_base(t, b, p) (b)
279
280#endif /* !CONFIG_SMP */
281
282/*
283 * Functions for the union type storage format of ktime_t which are
284 * too large for inlining:
285 */
286#if BITS_PER_LONG < 64
287/*
288 * Divide a ktime value by a nanosecond value
289 */
290s64 __ktime_divns(const ktime_t kt, s64 div)
291{
292 int sft = 0;
293 s64 dclc;
294 u64 tmp;
295
296 dclc = ktime_to_ns(kt);
297 tmp = dclc < 0 ? -dclc : dclc;
298
299 /* Make sure the divisor is less than 2^32: */
300 while (div >> 32) {
301 sft++;
302 div >>= 1;
303 }
304 tmp >>= sft;
305 do_div(tmp, (unsigned long) div);
306 return dclc < 0 ? -tmp : tmp;
307}
308EXPORT_SYMBOL_GPL(__ktime_divns);
309#endif /* BITS_PER_LONG >= 64 */
310
311/*
312 * Add two ktime values and do a safety check for overflow:
313 */
314ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
315{
316 ktime_t res = ktime_add_unsafe(lhs, rhs);
317
318 /*
319 * We use KTIME_SEC_MAX here, the maximum timeout which we can
320 * return to user space in a timespec:
321 */
322 if (res < 0 || res < lhs || res < rhs)
323 res = ktime_set(KTIME_SEC_MAX, 0);
324
325 return res;
326}
327
328EXPORT_SYMBOL_GPL(ktime_add_safe);
329
330#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
331
332static struct debug_obj_descr hrtimer_debug_descr;
333
334static void *hrtimer_debug_hint(void *addr)
335{
336 return ((struct hrtimer *) addr)->function;
337}
338
339/*
340 * fixup_init is called when:
341 * - an active object is initialized
342 */
343static bool hrtimer_fixup_init(void *addr, enum debug_obj_state state)
344{
345 struct hrtimer *timer = addr;
346
347 switch (state) {
348 case ODEBUG_STATE_ACTIVE:
349 hrtimer_cancel(timer);
350 debug_object_init(timer, &hrtimer_debug_descr);
351 return true;
352 default:
353 return false;
354 }
355}
356
357/*
358 * fixup_activate is called when:
359 * - an active object is activated
360 * - an unknown non-static object is activated
361 */
362static bool hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
363{
364 switch (state) {
365 case ODEBUG_STATE_ACTIVE:
366 WARN_ON(1);
367 /* fall through */
368 default:
369 return false;
370 }
371}
372
373/*
374 * fixup_free is called when:
375 * - an active object is freed
376 */
377static bool hrtimer_fixup_free(void *addr, enum debug_obj_state state)
378{
379 struct hrtimer *timer = addr;
380
381 switch (state) {
382 case ODEBUG_STATE_ACTIVE:
383 hrtimer_cancel(timer);
384 debug_object_free(timer, &hrtimer_debug_descr);
385 return true;
386 default:
387 return false;
388 }
389}
390
391static struct debug_obj_descr hrtimer_debug_descr = {
392 .name = "hrtimer",
393 .debug_hint = hrtimer_debug_hint,
394 .fixup_init = hrtimer_fixup_init,
395 .fixup_activate = hrtimer_fixup_activate,
396 .fixup_free = hrtimer_fixup_free,
397};
398
399static inline void debug_hrtimer_init(struct hrtimer *timer)
400{
401 debug_object_init(timer, &hrtimer_debug_descr);
402}
403
404static inline void debug_hrtimer_activate(struct hrtimer *timer,
405 enum hrtimer_mode mode)
406{
407 debug_object_activate(timer, &hrtimer_debug_descr);
408}
409
410static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
411{
412 debug_object_deactivate(timer, &hrtimer_debug_descr);
413}
414
415static inline void debug_hrtimer_free(struct hrtimer *timer)
416{
417 debug_object_free(timer, &hrtimer_debug_descr);
418}
419
420static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
421 enum hrtimer_mode mode);
422
423void hrtimer_init_on_stack(struct hrtimer *timer, clockid_t clock_id,
424 enum hrtimer_mode mode)
425{
426 debug_object_init_on_stack(timer, &hrtimer_debug_descr);
427 __hrtimer_init(timer, clock_id, mode);
428}
429EXPORT_SYMBOL_GPL(hrtimer_init_on_stack);
430
431void destroy_hrtimer_on_stack(struct hrtimer *timer)
432{
433 debug_object_free(timer, &hrtimer_debug_descr);
434}
435EXPORT_SYMBOL_GPL(destroy_hrtimer_on_stack);
436
437#else
438
439static inline void debug_hrtimer_init(struct hrtimer *timer) { }
440static inline void debug_hrtimer_activate(struct hrtimer *timer,
441 enum hrtimer_mode mode) { }
442static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
443#endif
444
445static inline void
446debug_init(struct hrtimer *timer, clockid_t clockid,
447 enum hrtimer_mode mode)
448{
449 debug_hrtimer_init(timer);
450 trace_hrtimer_init(timer, clockid, mode);
451}
452
453static inline void debug_activate(struct hrtimer *timer,
454 enum hrtimer_mode mode)
455{
456 debug_hrtimer_activate(timer, mode);
457 trace_hrtimer_start(timer, mode);
458}
459
460static inline void debug_deactivate(struct hrtimer *timer)
461{
462 debug_hrtimer_deactivate(timer);
463 trace_hrtimer_cancel(timer);
464}
465
466static struct hrtimer_clock_base *
467__next_base(struct hrtimer_cpu_base *cpu_base, unsigned int *active)
468{
469 unsigned int idx;
470
471 if (!*active)
472 return NULL;
473
474 idx = __ffs(*active);
475 *active &= ~(1U << idx);
476
477 return &cpu_base->clock_base[idx];
478}
479
480#define for_each_active_base(base, cpu_base, active) \
481 while ((base = __next_base((cpu_base), &(active))))
482
483static ktime_t __hrtimer_next_event_base(struct hrtimer_cpu_base *cpu_base,
484 const struct hrtimer *exclude,
485 unsigned int active,
486 ktime_t expires_next)
487{
488 struct hrtimer_clock_base *base;
489 ktime_t expires;
490
491 for_each_active_base(base, cpu_base, active) {
492 struct timerqueue_node *next;
493 struct hrtimer *timer;
494
495 next = timerqueue_getnext(&base->active);
496 timer = container_of(next, struct hrtimer, node);
497 if (timer == exclude) {
498 /* Get to the next timer in the queue. */
499 next = timerqueue_iterate_next(next);
500 if (!next)
501 continue;
502
503 timer = container_of(next, struct hrtimer, node);
504 }
505 expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
506 if (expires < expires_next) {
507 expires_next = expires;
508
509 /* Skip cpu_base update if a timer is being excluded. */
510 if (exclude)
511 continue;
512
513 if (timer->is_soft)
514 cpu_base->softirq_next_timer = timer;
515 else
516 cpu_base->next_timer = timer;
517 }
518 }
519 /*
520 * clock_was_set() might have changed base->offset of any of
521 * the clock bases so the result might be negative. Fix it up
522 * to prevent a false positive in clockevents_program_event().
523 */
524 if (expires_next < 0)
525 expires_next = 0;
526 return expires_next;
527}
528
529/*
530 * Recomputes cpu_base::*next_timer and returns the earliest expires_next but
531 * does not set cpu_base::*expires_next, that is done by hrtimer_reprogram.
532 *
533 * When a softirq is pending, we can ignore the HRTIMER_ACTIVE_SOFT bases,
534 * those timers will get run whenever the softirq gets handled, at the end of
535 * hrtimer_run_softirq(), hrtimer_update_softirq_timer() will re-add these bases.
536 *
537 * Therefore softirq values are those from the HRTIMER_ACTIVE_SOFT clock bases.
538 * The !softirq values are the minima across HRTIMER_ACTIVE_ALL, unless an actual
539 * softirq is pending, in which case they're the minima of HRTIMER_ACTIVE_HARD.
540 *
541 * @active_mask must be one of:
542 * - HRTIMER_ACTIVE_ALL,
543 * - HRTIMER_ACTIVE_SOFT, or
544 * - HRTIMER_ACTIVE_HARD.
545 */
546static ktime_t
547__hrtimer_get_next_event(struct hrtimer_cpu_base *cpu_base, unsigned int active_mask)
548{
549 unsigned int active;
550 struct hrtimer *next_timer = NULL;
551 ktime_t expires_next = KTIME_MAX;
552
553 if (!cpu_base->softirq_activated && (active_mask & HRTIMER_ACTIVE_SOFT)) {
554 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
555 cpu_base->softirq_next_timer = NULL;
556 expires_next = __hrtimer_next_event_base(cpu_base, NULL,
557 active, KTIME_MAX);
558
559 next_timer = cpu_base->softirq_next_timer;
560 }
561
562 if (active_mask & HRTIMER_ACTIVE_HARD) {
563 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
564 cpu_base->next_timer = next_timer;
565 expires_next = __hrtimer_next_event_base(cpu_base, NULL, active,
566 expires_next);
567 }
568
569 return expires_next;
570}
571
572static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
573{
574 ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
575 ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
576 ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
577
578 ktime_t now = ktime_get_update_offsets_now(&base->clock_was_set_seq,
579 offs_real, offs_boot, offs_tai);
580
581 base->clock_base[HRTIMER_BASE_REALTIME_SOFT].offset = *offs_real;
582 base->clock_base[HRTIMER_BASE_BOOTTIME_SOFT].offset = *offs_boot;
583 base->clock_base[HRTIMER_BASE_TAI_SOFT].offset = *offs_tai;
584
585 return now;
586}
587
588/*
589 * Is the high resolution mode active ?
590 */
591static inline int __hrtimer_hres_active(struct hrtimer_cpu_base *cpu_base)
592{
593 return IS_ENABLED(CONFIG_HIGH_RES_TIMERS) ?
594 cpu_base->hres_active : 0;
595}
596
597static inline int hrtimer_hres_active(void)
598{
599 return __hrtimer_hres_active(this_cpu_ptr(&hrtimer_bases));
600}
601
602/*
603 * Reprogram the event source with checking both queues for the
604 * next event
605 * Called with interrupts disabled and base->lock held
606 */
607static void
608hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
609{
610 ktime_t expires_next;
611
612 /*
613 * Find the current next expiration time.
614 */
615 expires_next = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_ALL);
616
617 if (cpu_base->next_timer && cpu_base->next_timer->is_soft) {
618 /*
619 * When the softirq is activated, hrtimer has to be
620 * programmed with the first hard hrtimer because soft
621 * timer interrupt could occur too late.
622 */
623 if (cpu_base->softirq_activated)
624 expires_next = __hrtimer_get_next_event(cpu_base,
625 HRTIMER_ACTIVE_HARD);
626 else
627 cpu_base->softirq_expires_next = expires_next;
628 }
629
630 if (skip_equal && expires_next == cpu_base->expires_next)
631 return;
632
633 cpu_base->expires_next = expires_next;
634
635 /*
636 * If hres is not active, hardware does not have to be
637 * reprogrammed yet.
638 *
639 * If a hang was detected in the last timer interrupt then we
640 * leave the hang delay active in the hardware. We want the
641 * system to make progress. That also prevents the following
642 * scenario:
643 * T1 expires 50ms from now
644 * T2 expires 5s from now
645 *
646 * T1 is removed, so this code is called and would reprogram
647 * the hardware to 5s from now. Any hrtimer_start after that
648 * will not reprogram the hardware due to hang_detected being
649 * set. So we'd effectivly block all timers until the T2 event
650 * fires.
651 */
652 if (!__hrtimer_hres_active(cpu_base) || cpu_base->hang_detected)
653 return;
654
655 tick_program_event(cpu_base->expires_next, 1);
656}
657
658/* High resolution timer related functions */
659#ifdef CONFIG_HIGH_RES_TIMERS
660
661/*
662 * High resolution timer enabled ?
663 */
664static bool hrtimer_hres_enabled __read_mostly = true;
665unsigned int hrtimer_resolution __read_mostly = LOW_RES_NSEC;
666EXPORT_SYMBOL_GPL(hrtimer_resolution);
667
668/*
669 * Enable / Disable high resolution mode
670 */
671static int __init setup_hrtimer_hres(char *str)
672{
673 return (kstrtobool(str, &hrtimer_hres_enabled) == 0);
674}
675
676__setup("highres=", setup_hrtimer_hres);
677
678/*
679 * hrtimer_high_res_enabled - query, if the highres mode is enabled
680 */
681static inline int hrtimer_is_hres_enabled(void)
682{
683 return hrtimer_hres_enabled;
684}
685
686/*
687 * Retrigger next event is called after clock was set
688 *
689 * Called with interrupts disabled via on_each_cpu()
690 */
691static void retrigger_next_event(void *arg)
692{
693 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
694
695 if (!__hrtimer_hres_active(base))
696 return;
697
698 raw_spin_lock(&base->lock);
699 hrtimer_update_base(base);
700 hrtimer_force_reprogram(base, 0);
701 raw_spin_unlock(&base->lock);
702}
703
704/*
705 * Switch to high resolution mode
706 */
707static void hrtimer_switch_to_hres(void)
708{
709 struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
710
711 if (tick_init_highres()) {
712 pr_warn("Could not switch to high resolution mode on CPU %u\n",
713 base->cpu);
714 return;
715 }
716 base->hres_active = 1;
717 hrtimer_resolution = HIGH_RES_NSEC;
718
719 tick_setup_sched_timer();
720 /* "Retrigger" the interrupt to get things going */
721 retrigger_next_event(NULL);
722}
723
724static void clock_was_set_work(struct work_struct *work)
725{
726 clock_was_set();
727}
728
729static DECLARE_WORK(hrtimer_work, clock_was_set_work);
730
731/*
732 * Called from timekeeping and resume code to reprogram the hrtimer
733 * interrupt device on all cpus.
734 */
735void clock_was_set_delayed(void)
736{
737 schedule_work(&hrtimer_work);
738}
739
740#else
741
742static inline int hrtimer_is_hres_enabled(void) { return 0; }
743static inline void hrtimer_switch_to_hres(void) { }
744static inline void retrigger_next_event(void *arg) { }
745
746#endif /* CONFIG_HIGH_RES_TIMERS */
747
748/*
749 * When a timer is enqueued and expires earlier than the already enqueued
750 * timers, we have to check, whether it expires earlier than the timer for
751 * which the clock event device was armed.
752 *
753 * Called with interrupts disabled and base->cpu_base.lock held
754 */
755static void hrtimer_reprogram(struct hrtimer *timer, bool reprogram)
756{
757 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
758 struct hrtimer_clock_base *base = timer->base;
759 ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
760
761 WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
762
763 /*
764 * CLOCK_REALTIME timer might be requested with an absolute
765 * expiry time which is less than base->offset. Set it to 0.
766 */
767 if (expires < 0)
768 expires = 0;
769
770 if (timer->is_soft) {
771 /*
772 * soft hrtimer could be started on a remote CPU. In this
773 * case softirq_expires_next needs to be updated on the
774 * remote CPU. The soft hrtimer will not expire before the
775 * first hard hrtimer on the remote CPU -
776 * hrtimer_check_target() prevents this case.
777 */
778 struct hrtimer_cpu_base *timer_cpu_base = base->cpu_base;
779
780 if (timer_cpu_base->softirq_activated)
781 return;
782
783 if (!ktime_before(expires, timer_cpu_base->softirq_expires_next))
784 return;
785
786 timer_cpu_base->softirq_next_timer = timer;
787 timer_cpu_base->softirq_expires_next = expires;
788
789 if (!ktime_before(expires, timer_cpu_base->expires_next) ||
790 !reprogram)
791 return;
792 }
793
794 /*
795 * If the timer is not on the current cpu, we cannot reprogram
796 * the other cpus clock event device.
797 */
798 if (base->cpu_base != cpu_base)
799 return;
800
801 /*
802 * If the hrtimer interrupt is running, then it will
803 * reevaluate the clock bases and reprogram the clock event
804 * device. The callbacks are always executed in hard interrupt
805 * context so we don't need an extra check for a running
806 * callback.
807 */
808 if (cpu_base->in_hrtirq)
809 return;
810
811 if (expires >= cpu_base->expires_next)
812 return;
813
814 /* Update the pointer to the next expiring timer */
815 cpu_base->next_timer = timer;
816 cpu_base->expires_next = expires;
817
818 /*
819 * If hres is not active, hardware does not have to be
820 * programmed yet.
821 *
822 * If a hang was detected in the last timer interrupt then we
823 * do not schedule a timer which is earlier than the expiry
824 * which we enforced in the hang detection. We want the system
825 * to make progress.
826 */
827 if (!__hrtimer_hres_active(cpu_base) || cpu_base->hang_detected)
828 return;
829
830 /*
831 * Program the timer hardware. We enforce the expiry for
832 * events which are already in the past.
833 */
834 tick_program_event(expires, 1);
835}
836
837/*
838 * Clock realtime was set
839 *
840 * Change the offset of the realtime clock vs. the monotonic
841 * clock.
842 *
843 * We might have to reprogram the high resolution timer interrupt. On
844 * SMP we call the architecture specific code to retrigger _all_ high
845 * resolution timer interrupts. On UP we just disable interrupts and
846 * call the high resolution interrupt code.
847 */
848void clock_was_set(void)
849{
850#ifdef CONFIG_HIGH_RES_TIMERS
851 /* Retrigger the CPU local events everywhere */
852 on_each_cpu(retrigger_next_event, NULL, 1);
853#endif
854 timerfd_clock_was_set();
855}
856
857/*
858 * During resume we might have to reprogram the high resolution timer
859 * interrupt on all online CPUs. However, all other CPUs will be
860 * stopped with IRQs interrupts disabled so the clock_was_set() call
861 * must be deferred.
862 */
863void hrtimers_resume(void)
864{
865 lockdep_assert_irqs_disabled();
866 /* Retrigger on the local CPU */
867 retrigger_next_event(NULL);
868 /* And schedule a retrigger for all others */
869 clock_was_set_delayed();
870}
871
872/*
873 * Counterpart to lock_hrtimer_base above:
874 */
875static inline
876void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
877{
878 raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
879}
880
881/**
882 * hrtimer_forward - forward the timer expiry
883 * @timer: hrtimer to forward
884 * @now: forward past this time
885 * @interval: the interval to forward
886 *
887 * Forward the timer expiry so it will expire in the future.
888 * Returns the number of overruns.
889 *
890 * Can be safely called from the callback function of @timer. If
891 * called from other contexts @timer must neither be enqueued nor
892 * running the callback and the caller needs to take care of
893 * serialization.
894 *
895 * Note: This only updates the timer expiry value and does not requeue
896 * the timer.
897 */
898u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
899{
900 u64 orun = 1;
901 ktime_t delta;
902
903 delta = ktime_sub(now, hrtimer_get_expires(timer));
904
905 if (delta < 0)
906 return 0;
907
908 if (WARN_ON(timer->state & HRTIMER_STATE_ENQUEUED))
909 return 0;
910
911 if (interval < hrtimer_resolution)
912 interval = hrtimer_resolution;
913
914 if (unlikely(delta >= interval)) {
915 s64 incr = ktime_to_ns(interval);
916
917 orun = ktime_divns(delta, incr);
918 hrtimer_add_expires_ns(timer, incr * orun);
919 if (hrtimer_get_expires_tv64(timer) > now)
920 return orun;
921 /*
922 * This (and the ktime_add() below) is the
923 * correction for exact:
924 */
925 orun++;
926 }
927 hrtimer_add_expires(timer, interval);
928
929 return orun;
930}
931EXPORT_SYMBOL_GPL(hrtimer_forward);
932
933/*
934 * enqueue_hrtimer - internal function to (re)start a timer
935 *
936 * The timer is inserted in expiry order. Insertion into the
937 * red black tree is O(log(n)). Must hold the base lock.
938 *
939 * Returns 1 when the new timer is the leftmost timer in the tree.
940 */
941static int enqueue_hrtimer(struct hrtimer *timer,
942 struct hrtimer_clock_base *base,
943 enum hrtimer_mode mode)
944{
945 debug_activate(timer, mode);
946
947 base->cpu_base->active_bases |= 1 << base->index;
948
949 timer->state = HRTIMER_STATE_ENQUEUED;
950
951 return timerqueue_add(&base->active, &timer->node);
952}
953
954/*
955 * __remove_hrtimer - internal function to remove a timer
956 *
957 * Caller must hold the base lock.
958 *
959 * High resolution timer mode reprograms the clock event device when the
960 * timer is the one which expires next. The caller can disable this by setting
961 * reprogram to zero. This is useful, when the context does a reprogramming
962 * anyway (e.g. timer interrupt)
963 */
964static void __remove_hrtimer(struct hrtimer *timer,
965 struct hrtimer_clock_base *base,
966 u8 newstate, int reprogram)
967{
968 struct hrtimer_cpu_base *cpu_base = base->cpu_base;
969 u8 state = timer->state;
970
971 timer->state = newstate;
972 if (!(state & HRTIMER_STATE_ENQUEUED))
973 return;
974
975 if (!timerqueue_del(&base->active, &timer->node))
976 cpu_base->active_bases &= ~(1 << base->index);
977
978 /*
979 * Note: If reprogram is false we do not update
980 * cpu_base->next_timer. This happens when we remove the first
981 * timer on a remote cpu. No harm as we never dereference
982 * cpu_base->next_timer. So the worst thing what can happen is
983 * an superflous call to hrtimer_force_reprogram() on the
984 * remote cpu later on if the same timer gets enqueued again.
985 */
986 if (reprogram && timer == cpu_base->next_timer)
987 hrtimer_force_reprogram(cpu_base, 1);
988}
989
990/*
991 * remove hrtimer, called with base lock held
992 */
993static inline int
994remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base, bool restart)
995{
996 if (hrtimer_is_queued(timer)) {
997 u8 state = timer->state;
998 int reprogram;
999
1000 /*
1001 * Remove the timer and force reprogramming when high
1002 * resolution mode is active and the timer is on the current
1003 * CPU. If we remove a timer on another CPU, reprogramming is
1004 * skipped. The interrupt event on this CPU is fired and
1005 * reprogramming happens in the interrupt handler. This is a
1006 * rare case and less expensive than a smp call.
1007 */
1008 debug_deactivate(timer);
1009 reprogram = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1010
1011 if (!restart)
1012 state = HRTIMER_STATE_INACTIVE;
1013
1014 __remove_hrtimer(timer, base, state, reprogram);
1015 return 1;
1016 }
1017 return 0;
1018}
1019
1020static inline ktime_t hrtimer_update_lowres(struct hrtimer *timer, ktime_t tim,
1021 const enum hrtimer_mode mode)
1022{
1023#ifdef CONFIG_TIME_LOW_RES
1024 /*
1025 * CONFIG_TIME_LOW_RES indicates that the system has no way to return
1026 * granular time values. For relative timers we add hrtimer_resolution
1027 * (i.e. one jiffie) to prevent short timeouts.
1028 */
1029 timer->is_rel = mode & HRTIMER_MODE_REL;
1030 if (timer->is_rel)
1031 tim = ktime_add_safe(tim, hrtimer_resolution);
1032#endif
1033 return tim;
1034}
1035
1036static void
1037hrtimer_update_softirq_timer(struct hrtimer_cpu_base *cpu_base, bool reprogram)
1038{
1039 ktime_t expires;
1040
1041 /*
1042 * Find the next SOFT expiration.
1043 */
1044 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
1045
1046 /*
1047 * reprogramming needs to be triggered, even if the next soft
1048 * hrtimer expires at the same time than the next hard
1049 * hrtimer. cpu_base->softirq_expires_next needs to be updated!
1050 */
1051 if (expires == KTIME_MAX)
1052 return;
1053
1054 /*
1055 * cpu_base->*next_timer is recomputed by __hrtimer_get_next_event()
1056 * cpu_base->*expires_next is only set by hrtimer_reprogram()
1057 */
1058 hrtimer_reprogram(cpu_base->softirq_next_timer, reprogram);
1059}
1060
1061static int __hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1062 u64 delta_ns, const enum hrtimer_mode mode,
1063 struct hrtimer_clock_base *base)
1064{
1065 struct hrtimer_clock_base *new_base;
1066
1067 /* Remove an active timer from the queue: */
1068 remove_hrtimer(timer, base, true);
1069
1070 if (mode & HRTIMER_MODE_REL)
1071 tim = ktime_add_safe(tim, base->get_time());
1072
1073 tim = hrtimer_update_lowres(timer, tim, mode);
1074
1075 hrtimer_set_expires_range_ns(timer, tim, delta_ns);
1076
1077 /* Switch the timer base, if necessary: */
1078 new_base = switch_hrtimer_base(timer, base, mode & HRTIMER_MODE_PINNED);
1079
1080 return enqueue_hrtimer(timer, new_base, mode);
1081}
1082
1083/**
1084 * hrtimer_start_range_ns - (re)start an hrtimer
1085 * @timer: the timer to be added
1086 * @tim: expiry time
1087 * @delta_ns: "slack" range for the timer
1088 * @mode: timer mode: absolute (HRTIMER_MODE_ABS) or
1089 * relative (HRTIMER_MODE_REL), and pinned (HRTIMER_MODE_PINNED);
1090 * softirq based mode is considered for debug purpose only!
1091 */
1092void hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1093 u64 delta_ns, const enum hrtimer_mode mode)
1094{
1095 struct hrtimer_clock_base *base;
1096 unsigned long flags;
1097
1098 /*
1099 * Check whether the HRTIMER_MODE_SOFT bit and hrtimer.is_soft
1100 * match.
1101 */
1102 WARN_ON_ONCE(!(mode & HRTIMER_MODE_SOFT) ^ !timer->is_soft);
1103
1104 base = lock_hrtimer_base(timer, &flags);
1105
1106 if (__hrtimer_start_range_ns(timer, tim, delta_ns, mode, base))
1107 hrtimer_reprogram(timer, true);
1108
1109 unlock_hrtimer_base(timer, &flags);
1110}
1111EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1112
1113/**
1114 * hrtimer_try_to_cancel - try to deactivate a timer
1115 * @timer: hrtimer to stop
1116 *
1117 * Returns:
1118 * 0 when the timer was not active
1119 * 1 when the timer was active
1120 * -1 when the timer is currently executing the callback function and
1121 * cannot be stopped
1122 */
1123int hrtimer_try_to_cancel(struct hrtimer *timer)
1124{
1125 struct hrtimer_clock_base *base;
1126 unsigned long flags;
1127 int ret = -1;
1128
1129 /*
1130 * Check lockless first. If the timer is not active (neither
1131 * enqueued nor running the callback, nothing to do here. The
1132 * base lock does not serialize against a concurrent enqueue,
1133 * so we can avoid taking it.
1134 */
1135 if (!hrtimer_active(timer))
1136 return 0;
1137
1138 base = lock_hrtimer_base(timer, &flags);
1139
1140 if (!hrtimer_callback_running(timer))
1141 ret = remove_hrtimer(timer, base, false);
1142
1143 unlock_hrtimer_base(timer, &flags);
1144
1145 return ret;
1146
1147}
1148EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1149
1150/**
1151 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1152 * @timer: the timer to be cancelled
1153 *
1154 * Returns:
1155 * 0 when the timer was not active
1156 * 1 when the timer was active
1157 */
1158int hrtimer_cancel(struct hrtimer *timer)
1159{
1160 for (;;) {
1161 int ret = hrtimer_try_to_cancel(timer);
1162
1163 if (ret >= 0)
1164 return ret;
1165 cpu_relax();
1166 }
1167}
1168EXPORT_SYMBOL_GPL(hrtimer_cancel);
1169
1170/**
1171 * hrtimer_get_remaining - get remaining time for the timer
1172 * @timer: the timer to read
1173 * @adjust: adjust relative timers when CONFIG_TIME_LOW_RES=y
1174 */
1175ktime_t __hrtimer_get_remaining(const struct hrtimer *timer, bool adjust)
1176{
1177 unsigned long flags;
1178 ktime_t rem;
1179
1180 lock_hrtimer_base(timer, &flags);
1181 if (IS_ENABLED(CONFIG_TIME_LOW_RES) && adjust)
1182 rem = hrtimer_expires_remaining_adjusted(timer);
1183 else
1184 rem = hrtimer_expires_remaining(timer);
1185 unlock_hrtimer_base(timer, &flags);
1186
1187 return rem;
1188}
1189EXPORT_SYMBOL_GPL(__hrtimer_get_remaining);
1190
1191#ifdef CONFIG_NO_HZ_COMMON
1192/**
1193 * hrtimer_get_next_event - get the time until next expiry event
1194 *
1195 * Returns the next expiry time or KTIME_MAX if no timer is pending.
1196 */
1197u64 hrtimer_get_next_event(void)
1198{
1199 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1200 u64 expires = KTIME_MAX;
1201 unsigned long flags;
1202
1203 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1204
1205 if (!__hrtimer_hres_active(cpu_base))
1206 expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_ALL);
1207
1208 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1209
1210 return expires;
1211}
1212
1213/**
1214 * hrtimer_next_event_without - time until next expiry event w/o one timer
1215 * @exclude: timer to exclude
1216 *
1217 * Returns the next expiry time over all timers except for the @exclude one or
1218 * KTIME_MAX if none of them is pending.
1219 */
1220u64 hrtimer_next_event_without(const struct hrtimer *exclude)
1221{
1222 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1223 u64 expires = KTIME_MAX;
1224 unsigned long flags;
1225
1226 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1227
1228 if (__hrtimer_hres_active(cpu_base)) {
1229 unsigned int active;
1230
1231 if (!cpu_base->softirq_activated) {
1232 active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
1233 expires = __hrtimer_next_event_base(cpu_base, exclude,
1234 active, KTIME_MAX);
1235 }
1236 active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
1237 expires = __hrtimer_next_event_base(cpu_base, exclude, active,
1238 expires);
1239 }
1240
1241 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1242
1243 return expires;
1244}
1245#endif
1246
1247static inline int hrtimer_clockid_to_base(clockid_t clock_id)
1248{
1249 if (likely(clock_id < MAX_CLOCKS)) {
1250 int base = hrtimer_clock_to_base_table[clock_id];
1251
1252 if (likely(base != HRTIMER_MAX_CLOCK_BASES))
1253 return base;
1254 }
1255 WARN(1, "Invalid clockid %d. Using MONOTONIC\n", clock_id);
1256 return HRTIMER_BASE_MONOTONIC;
1257}
1258
1259static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1260 enum hrtimer_mode mode)
1261{
1262 bool softtimer = !!(mode & HRTIMER_MODE_SOFT);
1263 int base = softtimer ? HRTIMER_MAX_CLOCK_BASES / 2 : 0;
1264 struct hrtimer_cpu_base *cpu_base;
1265
1266 memset(timer, 0, sizeof(struct hrtimer));
1267
1268 cpu_base = raw_cpu_ptr(&hrtimer_bases);
1269
1270 /*
1271 * POSIX magic: Relative CLOCK_REALTIME timers are not affected by
1272 * clock modifications, so they needs to become CLOCK_MONOTONIC to
1273 * ensure POSIX compliance.
1274 */
1275 if (clock_id == CLOCK_REALTIME && mode & HRTIMER_MODE_REL)
1276 clock_id = CLOCK_MONOTONIC;
1277
1278 base += hrtimer_clockid_to_base(clock_id);
1279 timer->is_soft = softtimer;
1280 timer->base = &cpu_base->clock_base[base];
1281 timerqueue_init(&timer->node);
1282}
1283
1284/**
1285 * hrtimer_init - initialize a timer to the given clock
1286 * @timer: the timer to be initialized
1287 * @clock_id: the clock to be used
1288 * @mode: The modes which are relevant for intitialization:
1289 * HRTIMER_MODE_ABS, HRTIMER_MODE_REL, HRTIMER_MODE_ABS_SOFT,
1290 * HRTIMER_MODE_REL_SOFT
1291 *
1292 * The PINNED variants of the above can be handed in,
1293 * but the PINNED bit is ignored as pinning happens
1294 * when the hrtimer is started
1295 */
1296void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1297 enum hrtimer_mode mode)
1298{
1299 debug_init(timer, clock_id, mode);
1300 __hrtimer_init(timer, clock_id, mode);
1301}
1302EXPORT_SYMBOL_GPL(hrtimer_init);
1303
1304/*
1305 * A timer is active, when it is enqueued into the rbtree or the
1306 * callback function is running or it's in the state of being migrated
1307 * to another cpu.
1308 *
1309 * It is important for this function to not return a false negative.
1310 */
1311bool hrtimer_active(const struct hrtimer *timer)
1312{
1313 struct hrtimer_clock_base *base;
1314 unsigned int seq;
1315
1316 do {
1317 base = READ_ONCE(timer->base);
1318 seq = raw_read_seqcount_begin(&base->seq);
1319
1320 if (timer->state != HRTIMER_STATE_INACTIVE ||
1321 base->running == timer)
1322 return true;
1323
1324 } while (read_seqcount_retry(&base->seq, seq) ||
1325 base != READ_ONCE(timer->base));
1326
1327 return false;
1328}
1329EXPORT_SYMBOL_GPL(hrtimer_active);
1330
1331/*
1332 * The write_seqcount_barrier()s in __run_hrtimer() split the thing into 3
1333 * distinct sections:
1334 *
1335 * - queued: the timer is queued
1336 * - callback: the timer is being ran
1337 * - post: the timer is inactive or (re)queued
1338 *
1339 * On the read side we ensure we observe timer->state and cpu_base->running
1340 * from the same section, if anything changed while we looked at it, we retry.
1341 * This includes timer->base changing because sequence numbers alone are
1342 * insufficient for that.
1343 *
1344 * The sequence numbers are required because otherwise we could still observe
1345 * a false negative if the read side got smeared over multiple consequtive
1346 * __run_hrtimer() invocations.
1347 */
1348
1349static void __run_hrtimer(struct hrtimer_cpu_base *cpu_base,
1350 struct hrtimer_clock_base *base,
1351 struct hrtimer *timer, ktime_t *now,
1352 unsigned long flags)
1353{
1354 enum hrtimer_restart (*fn)(struct hrtimer *);
1355 int restart;
1356
1357 lockdep_assert_held(&cpu_base->lock);
1358
1359 debug_deactivate(timer);
1360 base->running = timer;
1361
1362 /*
1363 * Separate the ->running assignment from the ->state assignment.
1364 *
1365 * As with a regular write barrier, this ensures the read side in
1366 * hrtimer_active() cannot observe base->running == NULL &&
1367 * timer->state == INACTIVE.
1368 */
1369 raw_write_seqcount_barrier(&base->seq);
1370
1371 __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
1372 fn = timer->function;
1373
1374 /*
1375 * Clear the 'is relative' flag for the TIME_LOW_RES case. If the
1376 * timer is restarted with a period then it becomes an absolute
1377 * timer. If its not restarted it does not matter.
1378 */
1379 if (IS_ENABLED(CONFIG_TIME_LOW_RES))
1380 timer->is_rel = false;
1381
1382 /*
1383 * The timer is marked as running in the CPU base, so it is
1384 * protected against migration to a different CPU even if the lock
1385 * is dropped.
1386 */
1387 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1388 trace_hrtimer_expire_entry(timer, now);
1389 restart = fn(timer);
1390 trace_hrtimer_expire_exit(timer);
1391 raw_spin_lock_irq(&cpu_base->lock);
1392
1393 /*
1394 * Note: We clear the running state after enqueue_hrtimer and
1395 * we do not reprogram the event hardware. Happens either in
1396 * hrtimer_start_range_ns() or in hrtimer_interrupt()
1397 *
1398 * Note: Because we dropped the cpu_base->lock above,
1399 * hrtimer_start_range_ns() can have popped in and enqueued the timer
1400 * for us already.
1401 */
1402 if (restart != HRTIMER_NORESTART &&
1403 !(timer->state & HRTIMER_STATE_ENQUEUED))
1404 enqueue_hrtimer(timer, base, HRTIMER_MODE_ABS);
1405
1406 /*
1407 * Separate the ->running assignment from the ->state assignment.
1408 *
1409 * As with a regular write barrier, this ensures the read side in
1410 * hrtimer_active() cannot observe base->running.timer == NULL &&
1411 * timer->state == INACTIVE.
1412 */
1413 raw_write_seqcount_barrier(&base->seq);
1414
1415 WARN_ON_ONCE(base->running != timer);
1416 base->running = NULL;
1417}
1418
1419static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now,
1420 unsigned long flags, unsigned int active_mask)
1421{
1422 struct hrtimer_clock_base *base;
1423 unsigned int active = cpu_base->active_bases & active_mask;
1424
1425 for_each_active_base(base, cpu_base, active) {
1426 struct timerqueue_node *node;
1427 ktime_t basenow;
1428
1429 basenow = ktime_add(now, base->offset);
1430
1431 while ((node = timerqueue_getnext(&base->active))) {
1432 struct hrtimer *timer;
1433
1434 timer = container_of(node, struct hrtimer, node);
1435
1436 /*
1437 * The immediate goal for using the softexpires is
1438 * minimizing wakeups, not running timers at the
1439 * earliest interrupt after their soft expiration.
1440 * This allows us to avoid using a Priority Search
1441 * Tree, which can answer a stabbing querry for
1442 * overlapping intervals and instead use the simple
1443 * BST we already have.
1444 * We don't add extra wakeups by delaying timers that
1445 * are right-of a not yet expired timer, because that
1446 * timer will have to trigger a wakeup anyway.
1447 */
1448 if (basenow < hrtimer_get_softexpires_tv64(timer))
1449 break;
1450
1451 __run_hrtimer(cpu_base, base, timer, &basenow, flags);
1452 }
1453 }
1454}
1455
1456static __latent_entropy void hrtimer_run_softirq(struct softirq_action *h)
1457{
1458 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1459 unsigned long flags;
1460 ktime_t now;
1461
1462 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1463
1464 now = hrtimer_update_base(cpu_base);
1465 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_SOFT);
1466
1467 cpu_base->softirq_activated = 0;
1468 hrtimer_update_softirq_timer(cpu_base, true);
1469
1470 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1471}
1472
1473#ifdef CONFIG_HIGH_RES_TIMERS
1474
1475/*
1476 * High resolution timer interrupt
1477 * Called with interrupts disabled
1478 */
1479void hrtimer_interrupt(struct clock_event_device *dev)
1480{
1481 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1482 ktime_t expires_next, now, entry_time, delta;
1483 unsigned long flags;
1484 int retries = 0;
1485
1486 BUG_ON(!cpu_base->hres_active);
1487 cpu_base->nr_events++;
1488 dev->next_event = KTIME_MAX;
1489
1490 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1491 entry_time = now = hrtimer_update_base(cpu_base);
1492retry:
1493 cpu_base->in_hrtirq = 1;
1494 /*
1495 * We set expires_next to KTIME_MAX here with cpu_base->lock
1496 * held to prevent that a timer is enqueued in our queue via
1497 * the migration code. This does not affect enqueueing of
1498 * timers which run their callback and need to be requeued on
1499 * this CPU.
1500 */
1501 cpu_base->expires_next = KTIME_MAX;
1502
1503 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1504 cpu_base->softirq_expires_next = KTIME_MAX;
1505 cpu_base->softirq_activated = 1;
1506 raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1507 }
1508
1509 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1510
1511 /* Reevaluate the clock bases for the next expiry */
1512 expires_next = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_ALL);
1513 /*
1514 * Store the new expiry value so the migration code can verify
1515 * against it.
1516 */
1517 cpu_base->expires_next = expires_next;
1518 cpu_base->in_hrtirq = 0;
1519 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1520
1521 /* Reprogramming necessary ? */
1522 if (!tick_program_event(expires_next, 0)) {
1523 cpu_base->hang_detected = 0;
1524 return;
1525 }
1526
1527 /*
1528 * The next timer was already expired due to:
1529 * - tracing
1530 * - long lasting callbacks
1531 * - being scheduled away when running in a VM
1532 *
1533 * We need to prevent that we loop forever in the hrtimer
1534 * interrupt routine. We give it 3 attempts to avoid
1535 * overreacting on some spurious event.
1536 *
1537 * Acquire base lock for updating the offsets and retrieving
1538 * the current time.
1539 */
1540 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1541 now = hrtimer_update_base(cpu_base);
1542 cpu_base->nr_retries++;
1543 if (++retries < 3)
1544 goto retry;
1545 /*
1546 * Give the system a chance to do something else than looping
1547 * here. We stored the entry time, so we know exactly how long
1548 * we spent here. We schedule the next event this amount of
1549 * time away.
1550 */
1551 cpu_base->nr_hangs++;
1552 cpu_base->hang_detected = 1;
1553 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1554
1555 delta = ktime_sub(now, entry_time);
1556 if ((unsigned int)delta > cpu_base->max_hang_time)
1557 cpu_base->max_hang_time = (unsigned int) delta;
1558 /*
1559 * Limit it to a sensible value as we enforce a longer
1560 * delay. Give the CPU at least 100ms to catch up.
1561 */
1562 if (delta > 100 * NSEC_PER_MSEC)
1563 expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1564 else
1565 expires_next = ktime_add(now, delta);
1566 tick_program_event(expires_next, 1);
1567 pr_warn_once("hrtimer: interrupt took %llu ns\n", ktime_to_ns(delta));
1568}
1569
1570/* called with interrupts disabled */
1571static inline void __hrtimer_peek_ahead_timers(void)
1572{
1573 struct tick_device *td;
1574
1575 if (!hrtimer_hres_active())
1576 return;
1577
1578 td = this_cpu_ptr(&tick_cpu_device);
1579 if (td && td->evtdev)
1580 hrtimer_interrupt(td->evtdev);
1581}
1582
1583#else /* CONFIG_HIGH_RES_TIMERS */
1584
1585static inline void __hrtimer_peek_ahead_timers(void) { }
1586
1587#endif /* !CONFIG_HIGH_RES_TIMERS */
1588
1589/*
1590 * Called from run_local_timers in hardirq context every jiffy
1591 */
1592void hrtimer_run_queues(void)
1593{
1594 struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1595 unsigned long flags;
1596 ktime_t now;
1597
1598 if (__hrtimer_hres_active(cpu_base))
1599 return;
1600
1601 /*
1602 * This _is_ ugly: We have to check periodically, whether we
1603 * can switch to highres and / or nohz mode. The clocksource
1604 * switch happens with xtime_lock held. Notification from
1605 * there only sets the check bit in the tick_oneshot code,
1606 * otherwise we might deadlock vs. xtime_lock.
1607 */
1608 if (tick_check_oneshot_change(!hrtimer_is_hres_enabled())) {
1609 hrtimer_switch_to_hres();
1610 return;
1611 }
1612
1613 raw_spin_lock_irqsave(&cpu_base->lock, flags);
1614 now = hrtimer_update_base(cpu_base);
1615
1616 if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1617 cpu_base->softirq_expires_next = KTIME_MAX;
1618 cpu_base->softirq_activated = 1;
1619 raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1620 }
1621
1622 __hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1623 raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1624}
1625
1626/*
1627 * Sleep related functions:
1628 */
1629static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
1630{
1631 struct hrtimer_sleeper *t =
1632 container_of(timer, struct hrtimer_sleeper, timer);
1633 struct task_struct *task = t->task;
1634
1635 t->task = NULL;
1636 if (task)
1637 wake_up_process(task);
1638
1639 return HRTIMER_NORESTART;
1640}
1641
1642void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, struct task_struct *task)
1643{
1644 sl->timer.function = hrtimer_wakeup;
1645 sl->task = task;
1646}
1647EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
1648
1649int nanosleep_copyout(struct restart_block *restart, struct timespec64 *ts)
1650{
1651 switch(restart->nanosleep.type) {
1652#ifdef CONFIG_COMPAT_32BIT_TIME
1653 case TT_COMPAT:
1654 if (put_old_timespec32(ts, restart->nanosleep.compat_rmtp))
1655 return -EFAULT;
1656 break;
1657#endif
1658 case TT_NATIVE:
1659 if (put_timespec64(ts, restart->nanosleep.rmtp))
1660 return -EFAULT;
1661 break;
1662 default:
1663 BUG();
1664 }
1665 return -ERESTART_RESTARTBLOCK;
1666}
1667
1668static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
1669{
1670 struct restart_block *restart;
1671
1672 hrtimer_init_sleeper(t, current);
1673
1674 do {
1675 set_current_state(TASK_INTERRUPTIBLE);
1676 hrtimer_start_expires(&t->timer, mode);
1677
1678 if (likely(t->task))
1679 freezable_schedule();
1680
1681 hrtimer_cancel(&t->timer);
1682 mode = HRTIMER_MODE_ABS;
1683
1684 } while (t->task && !signal_pending(current));
1685
1686 __set_current_state(TASK_RUNNING);
1687
1688 if (!t->task)
1689 return 0;
1690
1691 restart = &current->restart_block;
1692 if (restart->nanosleep.type != TT_NONE) {
1693 ktime_t rem = hrtimer_expires_remaining(&t->timer);
1694 struct timespec64 rmt;
1695
1696 if (rem <= 0)
1697 return 0;
1698 rmt = ktime_to_timespec64(rem);
1699
1700 return nanosleep_copyout(restart, &rmt);
1701 }
1702 return -ERESTART_RESTARTBLOCK;
1703}
1704
1705static long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
1706{
1707 struct hrtimer_sleeper t;
1708 int ret;
1709
1710 hrtimer_init_on_stack(&t.timer, restart->nanosleep.clockid,
1711 HRTIMER_MODE_ABS);
1712 hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
1713
1714 ret = do_nanosleep(&t, HRTIMER_MODE_ABS);
1715 destroy_hrtimer_on_stack(&t.timer);
1716 return ret;
1717}
1718
1719long hrtimer_nanosleep(const struct timespec64 *rqtp,
1720 const enum hrtimer_mode mode, const clockid_t clockid)
1721{
1722 struct restart_block *restart;
1723 struct hrtimer_sleeper t;
1724 int ret = 0;
1725 u64 slack;
1726
1727 slack = current->timer_slack_ns;
1728 if (dl_task(current) || rt_task(current))
1729 slack = 0;
1730
1731 hrtimer_init_on_stack(&t.timer, clockid, mode);
1732 hrtimer_set_expires_range_ns(&t.timer, timespec64_to_ktime(*rqtp), slack);
1733 ret = do_nanosleep(&t, mode);
1734 if (ret != -ERESTART_RESTARTBLOCK)
1735 goto out;
1736
1737 /* Absolute timers do not update the rmtp value and restart: */
1738 if (mode == HRTIMER_MODE_ABS) {
1739 ret = -ERESTARTNOHAND;
1740 goto out;
1741 }
1742
1743 restart = &current->restart_block;
1744 restart->fn = hrtimer_nanosleep_restart;
1745 restart->nanosleep.clockid = t.timer.base->clockid;
1746 restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
1747out:
1748 destroy_hrtimer_on_stack(&t.timer);
1749 return ret;
1750}
1751
1752#if !defined(CONFIG_64BIT_TIME) || defined(CONFIG_64BIT)
1753
1754SYSCALL_DEFINE2(nanosleep, struct __kernel_timespec __user *, rqtp,
1755 struct __kernel_timespec __user *, rmtp)
1756{
1757 struct timespec64 tu;
1758
1759 if (get_timespec64(&tu, rqtp))
1760 return -EFAULT;
1761
1762 if (!timespec64_valid(&tu))
1763 return -EINVAL;
1764
1765 current->restart_block.nanosleep.type = rmtp ? TT_NATIVE : TT_NONE;
1766 current->restart_block.nanosleep.rmtp = rmtp;
1767 return hrtimer_nanosleep(&tu, HRTIMER_MODE_REL, CLOCK_MONOTONIC);
1768}
1769
1770#endif
1771
1772#ifdef CONFIG_COMPAT_32BIT_TIME
1773
1774SYSCALL_DEFINE2(nanosleep_time32, struct old_timespec32 __user *, rqtp,
1775 struct old_timespec32 __user *, rmtp)
1776{
1777 struct timespec64 tu;
1778
1779 if (get_old_timespec32(&tu, rqtp))
1780 return -EFAULT;
1781
1782 if (!timespec64_valid(&tu))
1783 return -EINVAL;
1784
1785 current->restart_block.nanosleep.type = rmtp ? TT_COMPAT : TT_NONE;
1786 current->restart_block.nanosleep.compat_rmtp = rmtp;
1787 return hrtimer_nanosleep(&tu, HRTIMER_MODE_REL, CLOCK_MONOTONIC);
1788}
1789#endif
1790
1791/*
1792 * Functions related to boot-time initialization:
1793 */
1794int hrtimers_prepare_cpu(unsigned int cpu)
1795{
1796 struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
1797 int i;
1798
1799 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1800 cpu_base->clock_base[i].cpu_base = cpu_base;
1801 timerqueue_init_head(&cpu_base->clock_base[i].active);
1802 }
1803
1804 cpu_base->cpu = cpu;
1805 cpu_base->active_bases = 0;
1806 cpu_base->hres_active = 0;
1807 cpu_base->hang_detected = 0;
1808 cpu_base->next_timer = NULL;
1809 cpu_base->softirq_next_timer = NULL;
1810 cpu_base->expires_next = KTIME_MAX;
1811 cpu_base->softirq_expires_next = KTIME_MAX;
1812 return 0;
1813}
1814
1815#ifdef CONFIG_HOTPLUG_CPU
1816
1817static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
1818 struct hrtimer_clock_base *new_base)
1819{
1820 struct hrtimer *timer;
1821 struct timerqueue_node *node;
1822
1823 while ((node = timerqueue_getnext(&old_base->active))) {
1824 timer = container_of(node, struct hrtimer, node);
1825 BUG_ON(hrtimer_callback_running(timer));
1826 debug_deactivate(timer);
1827
1828 /*
1829 * Mark it as ENQUEUED not INACTIVE otherwise the
1830 * timer could be seen as !active and just vanish away
1831 * under us on another CPU
1832 */
1833 __remove_hrtimer(timer, old_base, HRTIMER_STATE_ENQUEUED, 0);
1834 timer->base = new_base;
1835 /*
1836 * Enqueue the timers on the new cpu. This does not
1837 * reprogram the event device in case the timer
1838 * expires before the earliest on this CPU, but we run
1839 * hrtimer_interrupt after we migrated everything to
1840 * sort out already expired timers and reprogram the
1841 * event device.
1842 */
1843 enqueue_hrtimer(timer, new_base, HRTIMER_MODE_ABS);
1844 }
1845}
1846
1847int hrtimers_dead_cpu(unsigned int scpu)
1848{
1849 struct hrtimer_cpu_base *old_base, *new_base;
1850 int i;
1851
1852 BUG_ON(cpu_online(scpu));
1853 tick_cancel_sched_timer(scpu);
1854
1855 /*
1856 * this BH disable ensures that raise_softirq_irqoff() does
1857 * not wakeup ksoftirqd (and acquire the pi-lock) while
1858 * holding the cpu_base lock
1859 */
1860 local_bh_disable();
1861 local_irq_disable();
1862 old_base = &per_cpu(hrtimer_bases, scpu);
1863 new_base = this_cpu_ptr(&hrtimer_bases);
1864 /*
1865 * The caller is globally serialized and nobody else
1866 * takes two locks at once, deadlock is not possible.
1867 */
1868 raw_spin_lock(&new_base->lock);
1869 raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
1870
1871 for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1872 migrate_hrtimer_list(&old_base->clock_base[i],
1873 &new_base->clock_base[i]);
1874 }
1875
1876 /*
1877 * The migration might have changed the first expiring softirq
1878 * timer on this CPU. Update it.
1879 */
1880 hrtimer_update_softirq_timer(new_base, false);
1881
1882 raw_spin_unlock(&old_base->lock);
1883 raw_spin_unlock(&new_base->lock);
1884
1885 /* Check, if we got expired work to do */
1886 __hrtimer_peek_ahead_timers();
1887 local_irq_enable();
1888 local_bh_enable();
1889 return 0;
1890}
1891
1892#endif /* CONFIG_HOTPLUG_CPU */
1893
1894void __init hrtimers_init(void)
1895{
1896 hrtimers_prepare_cpu(smp_processor_id());
1897 open_softirq(HRTIMER_SOFTIRQ, hrtimer_run_softirq);
1898}
1899
1900/**
1901 * schedule_hrtimeout_range_clock - sleep until timeout
1902 * @expires: timeout value (ktime_t)
1903 * @delta: slack in expires timeout (ktime_t)
1904 * @mode: timer mode
1905 * @clock_id: timer clock to be used
1906 */
1907int __sched
1908schedule_hrtimeout_range_clock(ktime_t *expires, u64 delta,
1909 const enum hrtimer_mode mode, clockid_t clock_id)
1910{
1911 struct hrtimer_sleeper t;
1912
1913 /*
1914 * Optimize when a zero timeout value is given. It does not
1915 * matter whether this is an absolute or a relative time.
1916 */
1917 if (expires && *expires == 0) {
1918 __set_current_state(TASK_RUNNING);
1919 return 0;
1920 }
1921
1922 /*
1923 * A NULL parameter means "infinite"
1924 */
1925 if (!expires) {
1926 schedule();
1927 return -EINTR;
1928 }
1929
1930 hrtimer_init_on_stack(&t.timer, clock_id, mode);
1931 hrtimer_set_expires_range_ns(&t.timer, *expires, delta);
1932
1933 hrtimer_init_sleeper(&t, current);
1934
1935 hrtimer_start_expires(&t.timer, mode);
1936
1937 if (likely(t.task))
1938 schedule();
1939
1940 hrtimer_cancel(&t.timer);
1941 destroy_hrtimer_on_stack(&t.timer);
1942
1943 __set_current_state(TASK_RUNNING);
1944
1945 return !t.task ? 0 : -EINTR;
1946}
1947
1948/**
1949 * schedule_hrtimeout_range - sleep until timeout
1950 * @expires: timeout value (ktime_t)
1951 * @delta: slack in expires timeout (ktime_t)
1952 * @mode: timer mode
1953 *
1954 * Make the current task sleep until the given expiry time has
1955 * elapsed. The routine will return immediately unless
1956 * the current task state has been set (see set_current_state()).
1957 *
1958 * The @delta argument gives the kernel the freedom to schedule the
1959 * actual wakeup to a time that is both power and performance friendly.
1960 * The kernel give the normal best effort behavior for "@expires+@delta",
1961 * but may decide to fire the timer earlier, but no earlier than @expires.
1962 *
1963 * You can set the task state as follows -
1964 *
1965 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
1966 * pass before the routine returns unless the current task is explicitly
1967 * woken up, (e.g. by wake_up_process()).
1968 *
1969 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
1970 * delivered to the current task or the current task is explicitly woken
1971 * up.
1972 *
1973 * The current task state is guaranteed to be TASK_RUNNING when this
1974 * routine returns.
1975 *
1976 * Returns 0 when the timer has expired. If the task was woken before the
1977 * timer expired by a signal (only possible in state TASK_INTERRUPTIBLE) or
1978 * by an explicit wakeup, it returns -EINTR.
1979 */
1980int __sched schedule_hrtimeout_range(ktime_t *expires, u64 delta,
1981 const enum hrtimer_mode mode)
1982{
1983 return schedule_hrtimeout_range_clock(expires, delta, mode,
1984 CLOCK_MONOTONIC);
1985}
1986EXPORT_SYMBOL_GPL(schedule_hrtimeout_range);
1987
1988/**
1989 * schedule_hrtimeout - sleep until timeout
1990 * @expires: timeout value (ktime_t)
1991 * @mode: timer mode
1992 *
1993 * Make the current task sleep until the given expiry time has
1994 * elapsed. The routine will return immediately unless
1995 * the current task state has been set (see set_current_state()).
1996 *
1997 * You can set the task state as follows -
1998 *
1999 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
2000 * pass before the routine returns unless the current task is explicitly
2001 * woken up, (e.g. by wake_up_process()).
2002 *
2003 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
2004 * delivered to the current task or the current task is explicitly woken
2005 * up.
2006 *
2007 * The current task state is guaranteed to be TASK_RUNNING when this
2008 * routine returns.
2009 *
2010 * Returns 0 when the timer has expired. If the task was woken before the
2011 * timer expired by a signal (only possible in state TASK_INTERRUPTIBLE) or
2012 * by an explicit wakeup, it returns -EINTR.
2013 */
2014int __sched schedule_hrtimeout(ktime_t *expires,
2015 const enum hrtimer_mode mode)
2016{
2017 return schedule_hrtimeout_range(expires, 0, mode);
2018}
2019EXPORT_SYMBOL_GPL(schedule_hrtimeout);